MiniMax claims Space Bunny and silently reroutes subscribers onto it the same day

On the afternoon of October 19, 2026, MiniMax's official account posted a short video captioned "Welcome back, Space Bunny" ¹². Space Bunny is the anonymous model that turned up in late September and reached number one in daily token volume on OpenRouter, a model marketplace, and on OpenCode, an open-source coding tool, within three days, holding the top three alongside DeepSeek's V4.1 Flash and GLM 5.3 Flash from Z.ai ¹. MiniMax has confirmed the bunny was an early build of M3.1 Flash, a stealth run that processed more than 200 trillion tokens↗ ³.

The switch no one's code asked for

The same date is stamped on a second, quieter document. From October 19, 2026, requests to M2.5 and M2.7 under MiniMax's M Plan and Token Plan subscriptions route automatically to M3.1 Flash; the old names keep working, allowances are unchanged, and both families stop entirely on November 18, 2026↗ ⁴. Nothing returns an error. The request still succeeds; the model answering it is one nobody chose. On the new model thinking cannot be switched off: a disabled flag returns HTTP 400 with error code 2013, an unset effort field defaults to max, and thinking tokens draw from the same allowance↗ ⁷. Request ceilings fall from 500 to 200 a minute ⁸.

A price row lands by afternoon

OrcaRouter, a model-routing service that resells API access, reported in the morning that MiniMax's commercial pages carried no per-token price for M3.1 Flash and no release note at all ⁸. Both were on the pages by the afternoon. The pricing table now lists M3.1 Flash at $0.15 input and $0.60 output per million tokens↗, half the $0.30 and $1.20 MiniMax charges for M3 up to 512k input, which doubles to $0.60 and $2.40 above it ⁵. A release note dated today calls it an ultra-efficient open-weight model↗ ⁶, with weights promised for October 23 and the license still undisclosed ³. Pay-as-you-go buyers get a halved rate. Subscribers get a hotter meter: OrcaRouter's arithmetic, offered strictly as arithmetic, puts a 40,000-request-a-day workload at roughly 2.9 billion extra thinking tokens a month on the default, about $3,400 at M3's old output rate ⁷.

The token math is the equity story

MiniMax Group, listed in Hong Kong as HKEX: 00100 ⁹, has turned itself into an API business. First-half 2026 revenue was $116.6 million, up 283.1 percent↗, and the Open Platform line, API calls plus subscriptions, grew 703.1 percent to $73.9 million, 63.4 percent of the total ⁹. Token consumption reached 20 times its January level by July, per co-founder and chief executive Yan Junjie ⁹. A migration that raises what each subscriber consumes against a fixed seat price lands squarely on that line. MiniMax's defense is real: the schedule is dated, pay-as-you-go users keep the old models until November 18 ⁴, and the per-token price halved ⁵. It does not answer why the most expensive effort setting is the one nobody has to choose.

What to watch

Weights land October 23, M2.5 and M2.7 stop answering on November 18, and every benchmark so far is vendor-reported, with no independent score attached ³⁴. No disclosure yet shows whether the max-effort default lifts what each subscriber burns. The number to watch is consumption per subscriber, and the next interim report is where it shows.